Market Research Agent: Governed Location Intelligence at Near-Zero Cost

A market research agent for a tong shui (Chinese dessert) shop, built so the running cost stays near zero. The trick is not a smarter model. It is refusing to let the model do the expensive things.

The questions a shop owner asks before opening are predictable: where is the foot traffic, how saturated is the area, what do competitors charge, what sells, what do reviews complain about. The tempting way to answer them is to wire a model to live SQL and a live Google Maps key. That answers everything, and it bills you on every single question.

So this agent never writes SQL. It picks from 10 fixed, parameterized queries over a 7-table data pack. Six of the ten are pure local SQL at near-zero cost. The other four read Google Maps results that were cached on a schedule, so Maps is billed only when the cache refreshes, never per question. The model's one job is to route the question to an approved tool and phrase the answer.

On the Klang Valley demo data it reads like a real brief:

• SS2 has the strongest evening foot traffic and 17 dessert shops already, so it is saturated
• Kepong carries the same evening crowd with only 4 shops mapped, the opportunity pocket
• competitors cluster around RM5.50, leaving the RM6.50 to RM7.50 band empty
• mango pomelo sago is the growing bestseller while plain green bean soup is slipping

The discipline a fixed toolset imposes is the point. A governed agent cannot leak data, cannot run a destructive query, and cannot quietly run up an API bill, because those capabilities were never built. The same 7 tables and 10 questions work for a chicken rice shop or an optician; only the data pack changes.

For a small business question, the useful design is not "can AI answer this?" It is "what is it allowed to do, and what does each answer cost?"